misuse of __AVX2__ etc., likely to cause miscompilation with GCC14

[mixal-iirec]

singleheader/simdjson.cpp contains:

#define SIMDJSON_CAN_ALWAYS_RUN_HASWELL ((SIMDJSON_IMPLEMENTATION_HASWELL) && (SIMDJSON_IS_X86_64) && (__AVX2__) && (__BMI__) && (__PCLMUL__) && (__LZCNT__))

/* ... */

#if !SIMDJSON_CAN_ALWAYS_RUN_HASWELL
SIMDJSON_TARGET_REGION("avx2,bmi,pclmul,lzcnt,popcnt")
#endif

/* Code to be compiled with haswell instruction sets. */

#if !SIMDJSON_CAN_ALWAYS_RUN_HASWELL
SIMDJSON_UNTARGET_REGION
#endif

Expanded:

#if ! ((SIMDJSON_IMPLEMENTATION_HASWELL) && (SIMDJSON_IS_X86_64) && (__AVX2__) && (__BMI__) && (__PCLMUL__) && (__LZCNT__))
_Pragma("GCC push_options")
_Pragma("GCC target(\"avx2,bmi,pclmul,lzcnt,popcnt\")")
#endif

/* Code to be compiled with haswell instruction sets. */

#if ! ((SIMDJSON_IMPLEMENTATION_HASWELL) && (SIMDJSON_IS_X86_64) && (__AVX2__) && (__BMI__) && (__PCLMUL__) && (__LZCNT__))
_Pragma("GCC pop_options")
#endif

Simplified:

#if ! __AVX2__
#pragma GCC push_options
#pragma GCC target("avx2")
#endif

/* Code to be compiled with haswell instruction sets. */

#if ! __AVX2__
#pragma GCC pop_options
#endif

GCC 13 and below defines __AVX2__ only when specified on command line.
This has been considered a bug, and was fixed for GCC14 https://gcc.gnu.org/bugzilla/show_bug.cgi?id=87299

In GCC 14, __AVX2__ is also defined with #pragma GCC target("avx2"). Thus the second __AVX2__ check will be always false and pop_options will never be called.
So all functions in the rest of the source file will be compiled with this instruction set added.

This is the underlying issue behind #2116

---

The fix should hopefully be straightforward (though I do not plan to fix it myself).
SIMDJSON_CAN_ALWAYS_RUN_HASWELL and any value guarding SIMDJSON_UNTARGET_REGION should not change in time:

#if ((SIMDJSON_IMPLEMENTATION_HASWELL) && (SIMDJSON_IS_X86_64) && (__AVX2__) && (__BMI__) && (__PCLMUL__) && (__LZCNT__))
#define SIMDJSON_CAN_ALWAYS_RUN_HASWELL 1
#else
#define SIMDJSON_CAN_ALWAYS_RUN_HASWELL 0
#endif

[lemire]

Thanks for the thorough analysis. We will apply your recommendation.

Note that GCC 14 is not released. Even so, I find that simdjson builds fine with GCC 14 under the current Fedora Rawhide:

$ gcc --version
gcc (GCC) 14.0.1 20240328 (Red Hat 14.0.1-0)
Copyright (C) 2024 Free Software Foundation, Inc.
This is free software; see the source for copying conditions.  There is NO
warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.

$ gcc -c simdjson.cpp  -std=c++17

I have removed the 'bug' marker. We only support released compilers.

Thanks again for the excellent analysis.

[mixal-iirec]

Note that GCC 14 is not released. Even so, I find that simdjson builds fine with GCC 14 under the current Fedora

Sorry that I haven't made it clear. Current version builds which is the bad case.
#2116 only made it easier to find.

But the resulting binary is wrong, for example after

gcc -c simdjson.cpp -O2 -std=c++17
objdump -S simdjson.o

<_ZNK8simdjson8westmere14implementation13validate_utf8EPKcm>
contains vpternlogq which is avx512 instruction, unsuported by westmere.

[lemire]

Thanks. The issue should be solved in any case in the current release. If not, just open a new pull request.